Classification Network-Guided Weighted K-Means Clustering for Multitouch Detection

In mobile devices, multitouch recognition technology is utilized to interact with the touch screen. Recognizing the coordinates of finger touches is the most important task in multitouch recognition. Several studies have been conducted to detect touch coordinates. However, these studies have been unable to detect exact coordinates when the big-finger problem occurs. The big-finger problem is a phenomenon in which a hole is generated in the center of the touch area when a large object approaches the touch screen, and it becomes more serious in a low-ground-mass environment. To solve the big-finger problem, we propose a novel method that detects touch coordinates using the clustering algorithm. We perform clustering on the touch data by treating the pixels of the touch data as data points. However, simply using the conventional clustering methods results in biased coordinate detection when touches of different sizes occur. Therefore, we propose the pixel intensity-weighted K-means clustering (PIKC) to prevent detection of biased coordinates. PIKC can detect unbiased coordinates by allocating touch pixel intensity values as weights in the centroid update process of clustering. Moreover, we accurately classify the number of touches using an efficient CNN-based multitouch classification network on adjacent-touch data. The predicted number of touches is used to determine the number of clusters for coordinate detection to further enhance the performance of our algorithm in adjacent-touch cases. Experiments demonstrate that our algorithm detects the touch coordinate most accurately compared to other algorithms in all touch scenarios.

Automated summary

In this study, we propose a novel method that uses clustering algorithm to detect multitouch coordinates, aiming to solve the big-finger problem. The method utilizes pixel intensity-weighted K-means clustering to prevent biased coordinate detection and employs an efficient CNN-based multitouch classification network to accurately classify the number of touches in adjacent-touch data. Experimental results demonstrate that our algorithm achieves the most accurate touch coordinate detection among all touch scenarios.